Confidence-Aware Transcription of Museum Specimen Labels with a LoRA-Tuned OCR Vision-Language Model

Millions of preserved museum specimens record their collection dates and localities only on physical labels that are handwritten, machine printed, faded, or written in historical Danish orthography, leaving key biodiversity metadata undigitized. The MuseumSCAT challenge formalizes the task: given a high-resolution specimen photograph (8192 x 5464 px), transcribe the collection date and locality exactly as written, or output MISSING, each with a confidence in [0, 1], with only 200 annotated training images (3,300 at test time). We adapt HunyuanOCR-1.5, a 1.0B-parameter end-to-end OCR vision-language model whose native-resolution encoder, handwriting and multilingual coverage, and verbatim "seen-text" training suit this regime, and fine-tune it with LoRA: rank-16 adapters in all seven decoder projections, vision tower frozen, 9.2M trainable parameters (0.92% of the model), with the loss masked to the JSON answer tokens. Per-field confidence is the exponential of the mean token log-probability, matching the competition's selective-prediction score (mean AURC over both fields). On a 21-image held-out split, fine-tuning on 179 images cuts the competition score from 0.955 to 0.495, lifts exact match from zero to 28.6% (dates) and 14.3% (localities), and yields confidences that separate MISSING calls (0.45 to 0.75) from transcribed values (0.24 to 0.58). In contrast to zero-shot use of the state-of-the-art base model, which transcribes nothing exactly, this shows that a small LoRA on under 200 labels converts a general lightweight OCR expert into a confidence-aware specimen-label transcriber.